Semantic search
- Context
- Standard term
Semantic search finds results by meaning. It gives your query and each document a list of numbers standing for what they say, then compares those lists, so related wording matches even when the exact words differ.
Each list of numbers, called an “embedding,” comes from a model trained on lots of text, which learned to give similar meanings similar numbers. The search returns the documents whose numbers sit closest to the query's.
Where you’ll meet it
Your email app adds AI search. You type "how much did we pay to fix the sink" and the plumber's invoice comes up first, though it only says "kitchen faucet repair." Right below it is the receipt for a bathroom sink you bought online.
What to do
Describe what you want in plain words when you don't remember the email's wording, and open the top result to confirm it answers your question. When you know the sender or the month, use the app's filters for them, since meaning-based search can blur exact names and numbers.
Easy to mix up
Keyword search can show why a result came up, since the matching words are right there. Semantic search gives a closeness score with no reason attached, so the why is harder to see.
Check yourself
A traveler searches a hotel chain's help pages for "which hotels do not allow pets?" Semantic search returns the page listing pet-friendly hotels. What went wrong?
Go deeperMany search systems combine meaning-based and keyword ranking
Keyword methods such as BM25 score documents by the query words they contain, giving rare words more weight. Hybrid search merges the keyword and semantic result lists, and a reranker, a second model reading the query and each candidate together, can then reorder the top results. Retrieval-augmented generation (RAG) systems, which look up passages and hand them to an AI model to answer from, often work this way.
An embedding model learns language from its training text, so new product names or a team's private jargon can land far from what they mean. To choose among semantic, keyword, and hybrid search for a collection, teams gather queries people typed, mark the right document for each, and score all three on that same set.